A steel structure guided wave detection signal generation method and system based on a diffusion probability model

Abstract

The application provides a steel structure guided wave detection signal generation method and system based on a diffusion probability model, and belongs to the field of guided wave signal data generation. The method collects guided wave signal data of a steel plate and a concrete-filled steel tube structure (including finite element simulation and actual measurement, defects and normal states), obtains a time-frequency graph through wavelet transformation, and inputs a U-Net diffusion model combined with multi-scale convolution and an attention mechanism for training and generation. The model introduces a wavelet feature decomposition and fusion module in the encoder, decoder and bottleneck layer to realize multi-scale enhancement of defect features. High-quality wavelet time-frequency graphs are generated through forward diffusion and reverse denoising, and the guided wave signals are restored through inverse wavelet transformation. The system includes a processor, a memory and a program, and sets data acquisition, wavelet transformation, diffusion model and inverse wavelet modules. The application can generate diversified high-fidelity guided wave signals, alleviate the shortage of defect samples, improve detection sensitivity and accuracy, and provide reliable data support for steel structure health monitoring.

Description

Technical Field

[0001] This invention relates to the field of guided wave signal data generation, and in particular to a method and system for generating guided wave detection signals for steel structures based on a diffusion probability model. Background Technology

[0002] Steel structures, due to their high strength and excellent durability, are widely used in major engineering projects such as bridges, factories, and high-rise buildings. With increasing service life, steel structures often face various damage problems caused by fatigue, corrosion, and welding defects. Timely and accurate detection of structural defects is crucial for ensuring project safety. Guided wave nondestructive testing technology, with its advantages of enabling rapid, wide-area structural inspection and high sensitivity to multiple types of damage, is playing an increasingly important role in the health monitoring of steel structures.

[0003] However, guided wave signals are affected by many factors such as structural materials, geometry, damage type, and environment. In actual detection, it is difficult to obtain high-quality and diverse defect samples, and the number of signals obtained from manual experiments and physical simulations is limited, making it difficult to meet the needs of intelligent detection algorithms for massive and complex samples. Traditional data augmentation or simulation methods often fail to fully reflect the distribution characteristics of real signals and complex noise environments, affecting the accuracy and generalization ability of the model.

[0004] Therefore, studying the generation method of guided wave signals has important theoretical significance and engineering value. This paper proposes a method and system for generating guided wave detection signals for steel structures based on the diffusion probability model. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for generating guided wave detection signals for steel structures based on a diffusion probability model, addressing the technical problem that existing guided wave detection signal generation methods for steel structures cannot meet the demands of intelligent detection algorithms for massive and complex samples. This method uses the U-Net model as the core architecture of the diffusion probability model. First, the original guided wave signal is converted into a wavelet time-frequency map, and then the diffusion probability model is used for high-quality image generation. Wavelet feature decomposition and fusion modules are introduced into each layer of the U-Net encoder and decoder, achieving multi-scale, fine-grained enhancement of defect features, thereby generating a high-fidelity wavelet time-frequency map. Finally, the guided wave signal is restored through inverse wavelet transform. This method not only significantly improves the expressive power of defect features and the quality of signal generation, but also provides efficient and accurate data support for steel structure health monitoring and non-destructive testing.

[0006] On the one hand, by generating diverse and high-quality guided wave signal data, the scarcity of actual defect samples and the difficulty of data collection can be effectively alleviated, establishing a more comprehensive and abundant data foundation for the field of steel structure health monitoring. On the other hand, the generated guided wave signals can cover different working conditions, material types, defect morphologies, and environmental factors, helping to improve the adaptability and generalization level of detection algorithms to complex engineering scenarios. Furthermore, this method can provide sufficient training and testing data for defect identification and health assessment models based on machine learning and deep learning, accelerating the research and development of intelligent detection technologies and promoting the transformation of the nondestructive testing field from experience-driven to data-driven and intelligent approaches. This has a positive effect on ensuring the safe operation of engineering structures and promoting technological progress in the industry.

[0007] This method acquires guided wave signal data (including finite element simulation and actual measurement, defect and normal states) of steel plate and steel-tube concrete structures, obtains time-frequency maps through wavelet transform, and inputs them into a U-Net diffusion model combining multi-scale convolution and attention mechanisms for training and generation. The model incorporates wavelet feature decomposition and fusion modules in the encoder, decoder, and bottleneck layer to achieve multi-scale enhancement of defect features. High-quality wavelet time-frequency maps are generated through forward diffusion and reverse denoising, and then the guided wave signals are restored through inverse wavelet transform. The system includes a processor, memory, and program, and is configured with data acquisition, wavelet transform, diffusion model, and inverse wavelet modules. It can generate diverse high-fidelity guided wave signals, alleviating the scarcity of defect samples, improving detection sensitivity and accuracy, and providing reliable data support for steel structure health monitoring.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for generating guided wave detection signals for steel structures based on a diffusion probability model, the method comprising the following steps:

[0010] S1, collect guided wave signal data of steel structure. The guided wave signal data includes defect data and normal data, and the data sources include simulation data and experimental data.

[0011] S2, perform wavelet transform on the guided wave signal data to generate a wavelet time-frequency diagram sequence corresponding to the guided wave signal, thus forming a wavelet time-frequency diagram;

[0012] S3. Establish and train a diffusion probability model with an attention mechanism. The diffusion probability model includes a forward diffusion process and a reverse denoising generation process. The forward diffusion process randomizes the signal features by gradually adding noise. The reverse denoising generation process uses the trained neural network to gradually reconstruct the signal. Wavelet features are extracted and fused by several scale convolutional layers and several head attention modules in the diffusion probability model to ensure that defective features are fully preserved and enhanced.

[0013] S4. Input the generated wavelet time-frequency plot into the trained diffusion probability model to generate a new wavelet time-frequency plot;

[0014] S5. Perform inverse wavelet transform on the generated new wavelet time-frequency diagram to obtain the final guided wave signal.

[0015] Furthermore, in step S3, the neural network structure of the diffusion probability model is a U-Net architecture, which includes an encoder, a bottleneck layer, and a decoder.

[0016] Furthermore, the encoder includes five downsampling modules, each consisting of a convolutional layer, an activation function layer, a normalization layer, a wavelet feature decomposition layer, a fusion module, and a pooling layer. The downsampling modules are used to gradually reduce the spatial resolution of the feature map and extract high- and low-frequency features at several scales.

[0017] Furthermore, the decoder includes five upsampling modules, each of which consists of an upsampling layer, a convolutional layer, an activation function layer, a normalization layer, a wavelet feature decomposition and fusion module. The upsampling modules are used to gradually restore the spatial resolution of the feature map, while fusing several scale features from the encoder.

[0018] The bottleneck layer consists of two convolutional layers, two activation function layers, two normalization layers, and one wavelet feature decomposition and fusion module. The bottleneck layer is used to further extract global features.

[0019] Furthermore, each layer of the encoder and decoder introduces a wavelet feature decomposition and fusion module to perform wavelet decomposition on the feature map, and then fuse the low-frequency and high-frequency subbands after convolution and attention processing, thereby achieving several scales and fine-grained enhancements to the defect features.

[0020] Furthermore, in step S3, the training data for the diffusion probability model includes simulation data obtained by the finite element method and experimental data obtained through actual steel structure waveguide experiments in the laboratory. The training data is classified according to normal and abnormal states, and is used to guide the model to learn the differences between defect features and normal features.

[0021] Furthermore, in step S3, the diffusion probability model employs the following loss function during training:

[0022] ;

[0023] in, , , , These are the weight coefficients of the loss function, and all are real numbers greater than 0. The loss is the mean square error of noise prediction. For structural similarity loss, For wavelet feature loss, The total variation loss is used to constrain the continuity of the generated wavelet time-frequency plot in the spatial dimension.

[0024] The noise prediction mean square error loss function is:

[0025] ;

[0026] in, Let represent the total number of training samples, and where i represents the i-th training sample. , The diffusion probability model represents the first... The noise obtained from the prediction of each sample This represents the actual noise corresponding to the i-th sample;

[0027] The structural similarity loss function is:

[0028] ;

[0029] in, Represents the structural similarity index. For model reconstruction output, These are real samples;

[0030] The wavelet feature loss function is:

[0031] ;

[0032] in, This represents the low-frequency component obtained after performing a wavelet transform on the input signal. This represents the high-frequency components obtained after performing a wavelet transform on the input signal. This represents the wavelet time-frequency plot of the generated sample. Wavelet time-frequency plot representing a real sample;

[0033] The total variation loss is:

[0034] ;

[0035] Where H represents the number of pixels in the wavelet time-frequency graph in the vertical direction, W represents the number of pixels in the wavelet time-frequency graph in the horizontal direction, and j and k represent the indices of the pixels in the vertical and horizontal directions, respectively. , , This indicates the location of the generated wavelet time-frequency plot. The pixel value at that location.

[0036] Furthermore, in step S5, after the guided wave signal is generated, the quality of the generated signal is initially evaluated by calculating the signal-to-noise ratio and the correlation coefficient with the reference signal. The higher the signal-to-noise ratio and the closer the correlation coefficient is to 1, the higher the quality of the generated signal.

[0037] A steel structure guided wave detection signal generation system based on a diffusion probability model includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it is configured to implement the above-mentioned signal generation method.

[0038] A system for generating guided wave detection signals for steel structures based on a diffusion probability model includes a data acquisition module, a wavelet transform module, a diffusion probability model processing module, and an inverse wavelet transform module. The data acquisition module is connected to the wavelet transform module, and both the wavelet transform and inverse wavelet transform modules are connected to the diffusion probability model processing module. The data acquisition module is used to acquire guided wave signal data of the steel structure, including defect data and normal data, and the data sources include simulated data and experimental data. The wavelet transform module is used to perform wavelet transform on the guided wave signal data to generate a wavelet time-frequency spectrum sequence corresponding to the guided wave signal, forming a wavelet time-frequency diagram. The diffusion probability model processing module... The module is used to build and train a diffusion probability model with an attention mechanism. The diffusion probability model includes a forward diffusion process and a reverse denoising generation process. The forward diffusion process randomizes the signal features by gradually adding noise, while the reverse denoising generation process gradually reconstructs the signal using the trained neural network. Wavelet features are extracted and fused in the diffusion probability model by several scale convolutional layers and several head attention modules to ensure that defective features are fully preserved and enhanced. The generated wavelet time-frequency map is input into the trained diffusion probability model to generate a new wavelet time-frequency map. The inverse wavelet transform module is used to perform inverse wavelet transform on the generated new wavelet time-frequency map to obtain the final guided wave signal.

[0039] The present invention, by adopting the above-described technical solution, has the following beneficial effects:

[0040] This invention enables multi-scale, fine-grained enhancement of defect features during guided wave signal generation, thereby generating high-quality, high-fidelity wavelet time-frequency maps and reconstructing realistic guided wave signals. This method not only significantly improves the sensitivity and accuracy of defect detection but also automatically and efficiently generates diverse guided wave signals, enriches data samples, enhances the generalization ability of the detection model, and reduces manual intervention and time costs. It possesses strong applicability and promotional value, providing a more efficient, intelligent, and reliable technical means for the field of non-destructive testing of steel structures. Attached Figure Description

[0041] Figure 1 This is a flowchart of the guided wave signal generation method of the present invention;

[0042] Figure 2 This is a schematic diagram of image generation using the diffusion probability model employed in this invention;

[0043] Figure 3 This is a schematic diagram of the U-Net network used in this invention;

[0044] Figure 4 This is a schematic diagram of the wavelet feature decomposition and fusion module in this invention;

[0045] Figure 5 This is a schematic diagram of the numerical simulation performed by the present invention to create a training set;

[0046] Figure 6 This is a partial time-domain graph of the training set in this invention;

[0047] Figure 7 This is a partial time-frequency diagram of the training set in this invention;

[0048] Figure 8 This is a time-domain plot of the guided wave signal generated by this invention;

[0049] Figure 9 This is a time-frequency diagram of the guided wave signal generated by this invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the invention, and these aspects of the invention can be implemented even without these specific details.

[0051] Example 1:

[0052] Taking concrete-filled steel tubular structures as an example, such as Figure 1-9 As shown, this invention provides a technical solution: a method for generating guided wave detection signals for steel structures based on a diffusion probability model, specifically executed according to S1 to S5 below to obtain the generated guided wave signals for steel-concrete composite structures:

[0053] S1, Collect guided wave signal data of steel structure, wherein the steel structure includes steel plate and steel pipe concrete structure, and the guided wave signal data includes defect data and normal data, and the data source includes simulation data and experimental data;

[0054] S2, perform wavelet transform on the guided wave signal data in step S1 to generate a wavelet time-frequency diagram sequence corresponding to the guided wave signal, thus forming a wavelet time-frequency diagram;

[0055] S3. Establish and train a diffusion probability model with an attention mechanism. The diffusion probability model includes a forward diffusion process and a reverse denoising generation process. The neural network structure is a U-Net architecture, including an encoder, a bottleneck layer, and a decoder. The forward diffusion process randomizes the signal features by gradually adding noise. The reverse denoising generation process uses the trained neural network to gradually reconstruct the signal. Wavelet features are extracted and fused by multi-scale convolutional layers and multi-head attention modules in this process, thereby ensuring that defective features are fully preserved and enhanced.

[0056] S4. Input the wavelet time-frequency plot generated in step S2 into the trained diffusion probability model to generate a new wavelet time-frequency plot.

[0057] S5. Perform inverse wavelet transform on the wavelet time-frequency diagram generated in step S4 to obtain the final guided wave signal.

[0058] The overall framework of the proposed method is as follows: Figure 1 As shown, in order to demonstrate specific embodiments of the present invention, the following were performed: Figure 5 The numerical simulation shown above collected the above signals to construct a training set. The signals in the training set and the wavelet time-frequency plot are shown below. Figure 6 , Figure 7 As shown in Table 1 below, the specific number of [items / items] in the training set is as follows.

[0059] Table 1 shows the details of the training set.

[0060] Data types quantity illustrate Numerical simulation 1,000 Covering the main operating conditions, the data is expanded to a sufficient number of training samples through slicing. Experimental data 80 Multiple batches of samples were collected for calibration and evaluation. total 1,080

[0061] As shown in Table 1, this invention combines numerical simulation data with experimental data when constructing the training set, with the amount of numerical simulation data significantly exceeding the amount of experimental data to meet the diffusion model's requirement for large-scale training samples. To ensure the consistency and sufficiency of the training data, the obtained time-series signals are preprocessed. Specifically, each time-series signal is segmented according to a preset window length (e.g., 256 to 1024 points), and an overlap rate of approximately 50% is set between adjacent windows to increase the number of effective samples and unify them to a size that matches the model input. When partitioning the dataset, the validation set and test set are reserved according to experimental batches or the entire working condition, respectively, to avoid sample leakage caused by data segmentation under the same working condition, thereby ensuring the objectivity and reliability of the model evaluation.

[0062] Figure 3 The U-Net network in this embodiment is shown. The encoder, wavelet feature decomposition and fusion module and decoder network in S3 will be described in detail below.

[0063] The specific structure of the encoder:

[0064] In this embodiment, the encoder part of the U-Net neural network structure includes five downsampling modules, denoted as Enc-1 to Enc-5. Each downsampling module is arranged in series in the following order: convolutional layer → activation function layer → normalization layer → wavelet feature decomposition and fusion module → pooling layer, used to gradually reduce the spatial resolution of the feature map and extract multi-scale high and low frequency features. The convolutional layer uses two-dimensional convolution, with a kernel size preferably of 3×3, a stride of 1, and padding of 1; the number of output channels increases with each layer. The activation function layer preferably uses ReLU; the normalization layer preferably uses BatchNorm2d. The wavelet feature decomposition and fusion module is used to perform wavelet decomposition and attention fusion processing on the input feature map. Specifically, it includes: decomposing the feature map into low-frequency and high-frequency subbands (such as LL, LH, HL, HH) in the low-resolution domain, performing channel mixing and attention weighting on each, upsampling back to the original scale, mapping back to the input channel number through a 1×1 convolution, and fusing it with the input using a residual method. The pooling layer preferably uses 2×2 max pooling with a step size of 2. Details are shown in Table 2.

[0065] Table 2 shows the encoder network architecture.

[0066] level Input dimensions Layer sequence Key parameters Output Channel Output size illustrate Enc-1 1 × 256 ×256 Conv→ ReLU→BN → Wavelet Fuse→MaxPool Conv:3×3,s=1,p=1; Pool:2×2,s=2 64 64 × 128 ×128 The wavelet module executes before pooling. Enc-2 64 × 128 ×128 Conv → ReLU →BN→ WaveletFuse→ MaxPool Conv:3×3,s=1,p=1; Pool:2×2,s=2 128 128 × 64 ×64 Maintain structural progression Enc-3 128 × 64 ×64 Conv → ReLU →BN→ WaveletFuse→ MaxPool Conv:3×3,s=1,p=1; Pool:2×2,s=2 256 256 × 32 ×32 Extracting mesoscale structures Enc-4 256 × 32 ×32 Conv → ReLU →BN→ WaveletFuse→ MaxPool Conv:3×3,s=1,p=1; Pool:2×2,s=2 512 512 × 16 ×16 Strengthening high-level semantics Enc-5 512 × 16 ×16 Conv → ReLU →BN→ WaveletFuse→ MaxPool Conv:3×3,s=1,p=1; Pool:2×2,s=2 1024 1024 × 8 ×8 Output to bottleneck layer

[0067] Wavelet eigenvalue decomposition and fusion module:

[0068] In this embodiment, the wavelet feature decomposition and fusion module is set at various levels of the encoder and decoder, and is used to perform wavelet domain multi-scale decomposition and fine-grained enhancement on the feature map. Its specific structure is as follows:

[0069] 1. Wavelet Decomposition Unit: A two-dimensional wavelet transform (preferably using the Haar wavelet kernel, but can also be replaced by orthogonal wavelet bases such as Daubechies) is employed to decompose the input feature map into a low-frequency subband LL and three high-frequency subbands LH, HL, and HH. This decomposition is implemented using a convolution with a stride of 2, thereby reducing the spatial resolution to half of the original while preserving the integrity of the information.

[0070] 2. Subband Processing Unit: Channel Mixing: Performs 1×1 convolution operations on the features of each subband obtained from the decomposition to achieve cross-channel information exchange and feature compression. Attention Weighting: Employs a channel attention mechanism (such as the SE module) to generate weight coefficients based on the global average features of the subbands, adaptively adjusting the importance of different subbands.

[0071] 3. Fusion Unit: Upsampling: The mixed and weighted sub-band features are restored to the spatial resolution of the original input through upsampling (e.g., bilinear interpolation). Feature Mapping: The fused multi-sub-band features are mapped back to the number of input channels using 1×1 convolution. Residual Connection: The above results are added element-wise to the module input to form a residual structure, thereby preserving the original information and enhancing the feature representation after wavelet domain enhancement.

[0072] Through the above structure, the wavelet feature decomposition and fusion module can enhance feature representation at both the frequency and spatial domains: the low-frequency subband is used to maintain the overall contour and structural information, while the high-frequency subband is used to highlight fine-grained defects and edge details, thereby significantly improving the model's ability to perceive complex defect features at multiple scales.

[0073] The specific structure of the decoder:

[0074] In this embodiment, the decoder part of the U-Net neural network structure includes five upsampling modules, numbered Dec-1 to Dec-5. Each upsampling module is connected in series in the following order: upsampling layer → convolutional layer → activation function layer → normalization layer → wavelet feature decomposition and fusion module, which is used to gradually restore the spatial resolution of the feature map and fuse multi-scale feature information from the encoder.

[0075] Specifically, the upsampling layer preferably uses bilinear interpolation with a scaling factor of 2; the convolutional layer uses two-dimensional convolution with a kernel size of 3×3, a stride of 1, padding of 1, and the number of output channels decreasing with each layer; the activation function layer preferably uses ReLU; the normalization layer preferably uses BatchNorm2d; the wavelet feature decomposition and fusion module has the same structure as the encoder module, performing wavelet decomposition and fusion processing on the upsampled features to enhance the fine-grained reconstruction capability in the decoding stage. Simultaneously, at the input of each decoder module, the encoder output features (skipped connections) of the corresponding layer are concatenated and fused with the upsampled features to maintain the combination of shallow details and deep semantics. Detailed information is shown in Table 3.

[0076] Table 3 shows the decoder network architecture.

[0077] level Input dimensions Input concatenation Layer sequence Key parameters Output Channel Output size illustrate December 1 1024 × 8 ×8 splicing with Enc-5 (1024) → Total 2048 channels Upsample →Conv → ReLU→ BN →Wavelet Fuse Upsample:×2; Conv:3×3,s=1,p=1 512 512 × 16 ×16 Recovery from the bottleneck layer, combined with Enc-5 features December 2 512 × 16 ×16 splicing with Enc-4 (512) → Total 1024 channels Upsample →Conv →ReLU→ BN →Wavelet Fuse Upsample:×2; Conv:3×3,s=1,p=1 256 256 × 32 ×32 Integrating high-level semantics of Enc-4 December 3 256 × 32 ×32 splicing with Enc-3 (256) → Total 512 channels Upsample →Conv → ReLU→ BN →Wavelet Fuse Upsample:×2; Conv:3×3,s=1,p=1 128 128 × 64 ×64 Integrating mid-layer features of Enc-3 December 4 128 × 64 ×64 splicing with Enc-2 (128) → Total 256 channels Upsample →Conv → ReLU→ BN →Wavelet Fuse Upsample:×2; Conv:3×3,s=1,p=1 64v 64 × 128 ×128 Shallow features of Enc-2 December 5 64 × 128 ×128 splicing with Enc-1 (64) → Total 128 channels Upsample →Conv →ReLU→ BN →Wavelet Fuse Upsample:×2; Conv:3×3,s=1,p=1 64 64 × 256 ×256 Restored to the original resolution, combined with the shallowest layer features

[0078] The present invention employs the following loss function during the training process:

[0079] ;

[0080] in, , , , These are the weight coefficients of the loss function, and all are real numbers greater than 0. The loss is the mean square error of noise prediction. For structural similarity loss, For wavelet feature loss, The total variation loss is used to constrain the continuity of the generated wavelet time-frequency plot in the spatial dimension.

[0081] The noise prediction mean square error loss function is:

[0082] ;

[0083] in, Let represent the total number of training samples, and where i represents the i-th training sample. , The diffusion probability model represents the first... The noise obtained from the prediction of each sample This represents the actual noise corresponding to the i-th sample;

[0084] The structural similarity loss function is:

[0085] ;

[0086] in, Represents the structural similarity index. For model reconstruction output, Using real samples, SSIM is an abbreviation for Structural Similarity Index Measure, a full-reference image quality assessment metric used to measure the similarity between two images. It compares images based on brightness, contrast, and structure, which better aligns with human visual perception of image quality and represents existing technological content.

[0087] The wavelet feature loss function is:

[0088] ;

[0089] in, This represents the low-frequency component obtained after performing a wavelet transform on the input signal. This represents the high-frequency components obtained after performing a wavelet transform on the input signal. This represents the wavelet time-frequency plot of the generated sample. Wavelet time-frequency plot representing a real sample;

[0090] The total variation loss is:

[0091] ;

[0092] Where H represents the number of pixels in the wavelet time-frequency graph in the vertical direction, W represents the number of pixels in the wavelet time-frequency graph in the horizontal direction, and j and k represent the indices of the pixels in the vertical and horizontal directions, respectively. , , This indicates the location of the generated wavelet time-frequency plot. The pixel value at that location.

[0093] In this embodiment, to verify the effectiveness of the method, the training hyperparameters of the model and the number of generated signals are set as shown in Table 4 below:

[0094] Table 4 shows the model hyperparameter and generated signal quantity settings.

[0095] category parameter Setting value illustrate Learning rate Learning Rate <![CDATA[3×10⁻ 4 ]]> AdamW optimizer, in conjunction with cosine annealing scheduling Batch size Batch Size 16 When video memory is insufficient, it can be reduced to 8. Training rounds Epochs 500 Early stop: No improvement after 20 rounds loss function Loss Function BCE + Dice (0.5:0.5) Simultaneously optimize pixel accuracy and region overlap. Regularization Regularization BatchNorm + Dropout(0.2) Stable training and prevention of overfitting Numerical simulation data Simulated Signals 1,000 records (expanded to ~20,000) Expand using a 50% overlap sliding window Experimental data Experimental Signals 80 items Multiple batches of data were collected for generalization constraints. Total Scale Training Samples ~20,080 Meets the training requirements of diffusion models

[0096] Under the above model structure and hyperparameter settings, a signal generation experiment was completed, generating 100 signals.

[0097] Example 2:

[0098] In this embodiment, a steel structure guided wave detection signal generation system based on a diffusion probability model is provided. The system includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it is configured to implement all the steps described in the foregoing method embodiments, including signal preprocessing, model training and inference, signal generation and post-processing, etc.

[0099] Example 3:

[0100] In this embodiment, a steel structure guided wave detection signal generation system based on a diffusion probability model is provided. The system includes a data acquisition module, a wavelet transform module, a diffusion probability model processing module, and an inverse wavelet transform module. The data acquisition module is connected to the wavelet transform module, and both the wavelet transform and inverse wavelet transform modules are connected to the diffusion probability model processing module. The data acquisition module is used to acquire guided wave signal data of the steel structure. The guided wave signal data includes defect data and normal data, and the data sources include simulated data and experimental data. The wavelet transform module is used to perform wavelet transform on the guided wave signal data to generate a wavelet time-frequency spectrum sequence corresponding to the guided wave signal, forming a wavelet time-frequency diagram. The probability model processing module is used to establish and train a diffusion probability model with an attention mechanism. The diffusion probability model includes a forward diffusion process and a reverse denoising generation process. The forward diffusion process randomizes the signal features by gradually adding noise, while the reverse denoising generation process gradually reconstructs the signal using the trained neural network. Wavelet features are extracted and fused in the diffusion probability model by several scale convolutional layers and several head attention modules to ensure that defective features are fully preserved and enhanced. The generated wavelet time-frequency map is input into the trained diffusion probability model to generate a new wavelet time-frequency map. The inverse wavelet transform module is used to perform inverse wavelet transform on the generated new wavelet time-frequency map to obtain the final guided wave signal.

[0101] The system comprises the following modules: a data acquisition module for acquiring guided wave signals from the steel structure under different operating conditions to obtain raw time-series signal data; a wavelet transform module for performing wavelet decomposition on the raw time-series signal to obtain multi-scale high- and low-frequency sub-band features; a diffusion probability model processing module for training and inference of the diffusion probability model in the wavelet time-frequency feature domain to generate and enhance the guided wave signal; and an inverse wavelet transform module for restoring the generated time-frequency features to the time-domain signal, thereby obtaining the final guided wave detection signal. Each module is executed collaboratively by the processor and memory. The computer program stored in memory is called during processor execution to implement the functional flow corresponding to the above modules.

[0102] Matters not covered in this invention are common knowledge.

[0103] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating guided wave detection signals for steel structures based on a diffusion probability model, characterized in that, The method includes the following steps: S1, collect guided wave signal data of steel structure. The guided wave signal data includes defect data and normal data, and the data sources include simulation data and experimental data. S2, perform wavelet transform on the guided wave signal data to generate a wavelet time-frequency diagram sequence corresponding to the guided wave signal, thus forming a wavelet time-frequency diagram; S3. Establish and train a diffusion probability model with an attention mechanism. The diffusion probability model includes a forward diffusion process and a reverse denoising generation process. The forward diffusion process randomizes the signal features by gradually adding noise. The reverse denoising generation process uses the trained neural network to gradually reconstruct the signal. Wavelet features are extracted and fused by several scale convolutional layers and several head attention modules in the diffusion probability model to ensure that defective features are fully preserved and enhanced. S4. Input the generated wavelet time-frequency plot into the trained diffusion probability model to generate a new wavelet time-frequency plot; S5, perform inverse wavelet transform on the generated new wavelet time-frequency diagram to obtain the final guided wave signal; In step S3, the neural network structure of the diffusion probability model is a U-Net architecture, which includes an encoder, a bottleneck layer, and a decoder. The encoder includes five downsampling modules. Each downsampling module consists of a convolutional layer, an activation function layer, a normalization layer, a wavelet feature decomposition and fusion module, and a pooling layer. The downsampling modules are used to gradually reduce the spatial resolution of the feature map and extract high and low frequency features at several scales. The decoder includes five upsampling modules. Each upsampling module consists of an upsampling layer, a convolutional layer, an activation function layer, a normalization layer, and a wavelet feature decomposition and fusion module. The upsampling module is used to gradually restore the spatial resolution of the feature map and fuse several high- and low-frequency features from the encoder. The bottleneck layer consists of two convolutional layers, two activation function layers, two normalization layers, and one wavelet feature decomposition and fusion module. The bottleneck layer is used to further extract global features. The wavelet feature decomposition and fusion module is used to perform wavelet decomposition and attention fusion processing on the input feature map. Specifically, it includes: decomposing the feature map into low-frequency and high-frequency sub-bands in the low-resolution domain, performing channel mixing and attention weighting on each sub-band, upsampling back to the original scale, mapping back to the input channel number through 1×1 convolution, and fusing it with the input feature map using a residual method.

2. The method for generating guided wave detection signals for steel structures based on a diffusion probability model according to claim 1, characterized in that: In step S3, the training data for the diffusion probability model includes simulation data obtained by the finite element method and experimental data obtained through actual steel structure waveguide experiments in the laboratory. The training data is classified according to normal and abnormal states, and is used to guide the model to learn the differences between defect features and normal features.

3. The method for generating guided wave detection signals for steel structures based on a diffusion probability model according to claim 1, characterized in that: In step S3, the diffusion probability model uses the following loss function during training: ,in, , , , These are the weight coefficients of the loss function, and all are real numbers greater than 0. The loss is the mean square error of noise prediction. For structural similarity loss, For wavelet feature loss, The total variation loss is used to constrain the continuity of the generated wavelet time-frequency plot in the spatial dimension. The mean square error loss for noise prediction is: ,in, Let represent the total number of training samples, and let i represent the i-th training sample. , This represents the noise predicted by the diffusion probability model for the i-th sample. This represents the actual noise corresponding to the i-th sample; The structural similarity loss is: ,in, Represents the structural similarity index. For model reconstruction output, These are real samples; The wavelet feature loss is: ,in, This represents the low-frequency component obtained after performing a wavelet transform on the input signal. This represents the high-frequency components obtained after performing a wavelet transform on the input signal. This represents the wavelet time-frequency plot of the generated sample. Wavelet time-frequency plot representing a real sample; The total variation loss is: Where H represents the number of pixels in the wavelet time-frequency graph in the vertical direction, W represents the number of pixels in the wavelet time-frequency graph in the horizontal direction, and j and k represent the indices of the pixels in the vertical and horizontal directions, respectively. , , This indicates the location of the generated wavelet time-frequency plot. The pixel value at that location.

4. The method for generating guided wave detection signals for steel structures based on a diffusion probability model according to claim 1, characterized in that: In step S5, after the guided wave signal is generated, the quality of the generated signal is initially evaluated by calculating the signal-to-noise ratio and the correlation coefficient with the reference signal. The higher the signal-to-noise ratio and the closer the correlation coefficient is to 1, the higher the quality of the generated signal.

5. A steel structure guided wave detection signal generation system based on a diffusion probability model, characterized in that: The method includes a processor, a memory, and a computer program stored in the memory, wherein the processor, when executing the computer program, is configured to implement the method as described in any one of claims 1 to 4.

6. A steel structure guided wave detection signal generation system based on a diffusion probability model, characterized in that: The system includes a data acquisition module, a wavelet transform module, a diffusion probability model processing module, and an inverse wavelet transform module. The data acquisition module is connected to the wavelet transform module, and both the wavelet transform and inverse wavelet transform modules are connected to the diffusion probability model processing module. The data acquisition module is used to acquire guided wave signal data of the steel structure. The guided wave signal data includes defect data and normal data, and the data sources include simulated data and experimental data. The wavelet transform module is used to perform wavelet transform on the guided wave signal data to generate the wavelet time-frequency spectrum sequence corresponding to the guided wave signal, forming a wavelet time-frequency diagram. The diffusion probability model processing module is used to establish and train attention-based... The diffusion probability model of the mechanism includes a forward diffusion process and a reverse denoising generation process. The forward diffusion process randomizes the signal features by gradually adding noise, while the reverse denoising generation process gradually reconstructs the signal using a trained neural network. Wavelet features are extracted and fused by several scale convolutional layers and several head attention modules in the diffusion probability model to ensure that defective features are fully preserved and enhanced. The generated wavelet time-frequency map is input into the trained diffusion probability model to generate a new wavelet time-frequency map. The inverse wavelet transform module is used to perform inverse wavelet transform on the generated new wavelet time-frequency map to obtain the final guided wave signal. The neural network structure of the diffusion probability model is the U-Net architecture, which includes an encoder, a bottleneck layer, and a decoder. The encoder includes five downsampling modules. Each downsampling module consists of a convolutional layer, an activation function layer, a normalization layer, a wavelet feature decomposition and fusion module, and a pooling layer. The downsampling modules are used to gradually reduce the spatial resolution of the feature map and extract high and low frequency features at several scales. The decoder includes five upsampling modules. Each upsampling module consists of an upsampling layer, a convolutional layer, an activation function layer, a normalization layer, and a wavelet feature decomposition and fusion module. The upsampling module is used to gradually restore the spatial resolution of the feature map and fuse several high- and low-frequency features from the encoder. The bottleneck layer consists of two convolutional layers, two activation function layers, two normalization layers, and one wavelet feature decomposition and fusion module. The bottleneck layer is used to further extract global features. The wavelet feature decomposition and fusion module is used to perform wavelet decomposition and attention fusion processing on the input feature map. Specifically, it includes: decomposing the feature map into low-frequency and high-frequency sub-bands in the low-resolution domain, performing channel mixing and attention weighting on each sub-band, upsampling back to the original scale, mapping back to the input channel number through 1×1 convolution, and fusing it with the input feature map using a residual method.

Images